1 | Inlining |
2 | ======== |
3 | |
4 | There are several options that control which calls the analyzer will consider for |
5 | inlining. The major one is ``-analyzer-config ipa``: |
6 | |
7 | * ``analyzer-config ipa=none`` - All inlining is disabled. This is the only mode |
8 | available in LLVM 3.1 and earlier and in Xcode 4.3 and earlier. |
9 | |
10 | * ``analyzer-config ipa=basic-inlining`` - Turns on inlining for C functions, C++ |
11 | static member functions, and blocks -- essentially, the calls that behave |
12 | like simple C function calls. This is essentially the mode used in |
13 | Xcode 4.4. |
14 | |
15 | * ``analyzer-config ipa=inlining`` - Turns on inlining when we can confidently find |
16 | the function/method body corresponding to the call. (C functions, static |
17 | functions, devirtualized C++ methods, Objective-C class methods, Objective-C |
18 | instance methods when ExprEngine is confident about the dynamic type of the |
19 | instance). |
20 | |
21 | * ``analyzer-config ipa=dynamic`` - Inline instance methods for which the type is |
22 | determined at runtime and we are not 100% sure that our type info is |
23 | correct. For virtual calls, inline the most plausible definition. |
24 | |
25 | * ``analyzer-config ipa=dynamic-bifurcate`` - Same as -analyzer-config ipa=dynamic, |
26 | but the path is split. We inline on one branch and do not inline on the |
27 | other. This mode does not drop the coverage in cases when the parent class |
28 | has code that is only exercised when some of its methods are overridden. |
29 | |
30 | Currently, ``-analyzer-config ipa=dynamic-bifurcate`` is the default mode. |
31 | |
32 | While ``-analyzer-config ipa`` determines in general how aggressively the analyzer |
33 | will try to inline functions, several additional options control which types of |
34 | functions can inlined, in an all-or-nothing way. These options use the |
35 | analyzer's configuration table, so they are all specified as follows: |
36 | |
37 | ``-analyzer-config OPTION=VALUE`` |
38 | |
39 | c++-inlining |
40 | ------------ |
41 | |
42 | This option controls which C++ member functions may be inlined. |
43 | |
44 | ``-analyzer-config c++-inlining=[none | methods | constructors | destructors]`` |
45 | |
46 | Each of these modes implies that all the previous member function kinds will be |
47 | inlined as well; it doesn't make sense to inline destructors without inlining |
48 | constructors, for example. |
49 | |
50 | The default c++-inlining mode is 'destructors', meaning that all member |
51 | functions with visible definitions will be considered for inlining. In some |
52 | cases the analyzer may still choose not to inline the function. |
53 | |
54 | Note that under 'constructors', constructors for types with non-trivial |
55 | destructors will not be inlined. Additionally, no C++ member functions will be |
56 | inlined under -analyzer-config ipa=none or -analyzer-config ipa=basic-inlining, |
57 | regardless of the setting of the c++-inlining mode. |
58 | |
59 | c++-template-inlining |
60 | ^^^^^^^^^^^^^^^^^^^^^ |
61 | |
62 | This option controls whether C++ templated functions may be inlined. |
63 | |
64 | ``-analyzer-config c++-template-inlining=[true | false]`` |
65 | |
66 | Currently, template functions are considered for inlining by default. |
67 | |
68 | The motivation behind this option is that very generic code can be a source |
69 | of false positives, either by considering paths that the caller considers |
70 | impossible (by some unstated precondition), or by inlining some but not all |
71 | of a deep implementation of a function. |
72 | |
73 | c++-stdlib-inlining |
74 | ^^^^^^^^^^^^^^^^^^^ |
75 | |
76 | This option controls whether functions from the C++ standard library, including |
77 | methods of the container classes in the Standard Template Library, should be |
78 | considered for inlining. |
79 | |
80 | ``-analyzer-config c++-stdlib-inlining=[true | false]`` |
81 | |
82 | Currently, C++ standard library functions are considered for inlining by |
83 | default. |
84 | |
85 | The standard library functions and the STL in particular are used ubiquitously |
86 | enough that our tolerance for false positives is even lower here. A false |
87 | positive due to poor modeling of the STL leads to a poor user experience, since |
88 | most users would not be comfortable adding assertions to system headers in order |
89 | to silence analyzer warnings. |
90 | |
91 | c++-container-inlining |
92 | ^^^^^^^^^^^^^^^^^^^^^^ |
93 | |
94 | This option controls whether constructors and destructors of "container" types |
95 | should be considered for inlining. |
96 | |
97 | ``-analyzer-config c++-container-inlining=[true | false]`` |
98 | |
99 | Currently, these constructors and destructors are NOT considered for inlining |
100 | by default. |
101 | |
102 | The current implementation of this setting checks whether a type has a member |
103 | named 'iterator' or a member named 'begin'; these names are idiomatic in C++, |
104 | with the latter specified in the C++11 standard. The analyzer currently does a |
105 | fairly poor job of modeling certain data structure invariants of container-like |
106 | objects. For example, these three expressions should be equivalent: |
107 | |
108 | |
109 | .. code-block:: cpp |
110 | |
111 | std::distance(c.begin(), c.end()) == 0 |
112 | c.begin() == c.end() |
113 | c.empty() |
114 | |
115 | Many of these issues are avoided if containers always have unknown, symbolic |
116 | state, which is what happens when their constructors are treated as opaque. |
117 | In the future, we may decide specific containers are "safe" to model through |
118 | inlining, or choose to model them directly using checkers instead. |
119 | |
120 | |
121 | Basics of Implementation |
122 | ------------------------ |
123 | |
124 | The low-level mechanism of inlining a function is handled in |
125 | ExprEngine::inlineCall and ExprEngine::processCallExit. |
126 | |
127 | If the conditions are right for inlining, a CallEnter node is created and added |
128 | to the analysis work list. The CallEnter node marks the change to a new |
129 | LocationContext representing the called function, and its state includes the |
130 | contents of the new stack frame. When the CallEnter node is actually processed, |
131 | its single successor will be a edge to the first CFG block in the function. |
132 | |
133 | Exiting an inlined function is a bit more work, fortunately broken up into |
134 | reasonable steps: |
135 | |
136 | 1. The CoreEngine realizes we're at the end of an inlined call and generates a |
137 | CallExitBegin node. |
138 | |
139 | 2. ExprEngine takes over (in processCallExit) and finds the return value of the |
140 | function, if it has one. This is bound to the expression that triggered the |
141 | call. (In the case of calls without origin expressions, such as destructors, |
142 | this step is skipped.) |
143 | |
144 | 3. Dead symbols and bindings are cleaned out from the state, including any local |
145 | bindings. |
146 | |
147 | 4. A CallExitEnd node is generated, which marks the transition back to the |
148 | caller's LocationContext. |
149 | |
150 | 5. Custom post-call checks are processed and the final nodes are pushed back |
151 | onto the work list, so that evaluation of the caller can continue. |
152 | |
153 | Retry Without Inlining |
154 | ^^^^^^^^^^^^^^^^^^^^^^ |
155 | |
156 | In some cases, we would like to retry analysis without inlining a particular |
157 | call. |
158 | |
159 | Currently, we use this technique to recover coverage in case we stop |
160 | analyzing a path due to exceeding the maximum block count inside an inlined |
161 | function. |
162 | |
163 | When this situation is detected, we walk up the path to find the first node |
164 | before inlining was started and enqueue it on the WorkList with a special |
165 | ReplayWithoutInlining bit added to it (ExprEngine::replayWithoutInlining). The |
166 | path is then re-analyzed from that point without inlining that particular call. |
167 | |
168 | Deciding When to Inline |
169 | ^^^^^^^^^^^^^^^^^^^^^^^ |
170 | |
171 | In general, the analyzer attempts to inline as much as possible, since it |
172 | provides a better summary of what actually happens in the program. There are |
173 | some cases, however, where the analyzer chooses not to inline: |
174 | |
175 | - If there is no definition available for the called function or method. In |
176 | this case, there is no opportunity to inline. |
177 | |
178 | - If the CFG cannot be constructed for a called function, or the liveness |
179 | cannot be computed. These are prerequisites for analyzing a function body, |
180 | with or without inlining. |
181 | |
182 | - If the LocationContext chain for a given ExplodedNode reaches a maximum cutoff |
183 | depth. This prevents unbounded analysis due to infinite recursion, but also |
184 | serves as a useful cutoff for performance reasons. |
185 | |
186 | - If the function is variadic. This is not a hard limitation, but an engineering |
187 | limitation. |
188 | |
189 | Tracked by: <rdar://problem/12147064> Support inlining of variadic functions |
190 | |
191 | - In C++, constructors are not inlined unless the destructor call will be |
192 | processed by the ExprEngine. Thus, if the CFG was built without nodes for |
193 | implicit destructors, or if the destructors for the given object are not |
194 | represented in the CFG, the constructor will not be inlined. (As an exception, |
195 | constructors for objects with trivial constructors can still be inlined.) |
196 | See "C++ Caveats" below. |
197 | |
198 | - In C++, ExprEngine does not inline custom implementations of operator 'new' |
199 | or operator 'delete', nor does it inline the constructors and destructors |
200 | associated with these. See "C++ Caveats" below. |
201 | |
202 | - Calls resulting in "dynamic dispatch" are specially handled. See more below. |
203 | |
204 | - The FunctionSummaries map stores additional information about declarations, |
205 | some of which is collected at runtime based on previous analyses. |
206 | We do not inline functions which were not profitable to inline in a different |
207 | context (for example, if the maximum block count was exceeded; see |
208 | "Retry Without Inlining"). |
209 | |
210 | |
211 | Dynamic Calls and Devirtualization |
212 | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ |
213 | |
214 | "Dynamic" calls are those that are resolved at runtime, such as C++ virtual |
215 | method calls and Objective-C message sends. Due to the path-sensitive nature of |
216 | the analysis, the analyzer may be able to reason about the dynamic type of the |
217 | object whose method is being called and thus "devirtualize" the call. |
218 | |
219 | This path-sensitive devirtualization occurs when the analyzer can determine what |
220 | method would actually be called at runtime. This is possible when the type |
221 | information is constrained enough for a simulated C++/Objective-C object that |
222 | the analyzer can make such a decision. |
223 | |
224 | DynamicTypeInfo |
225 | ^^^^^^^^^^^^^^^ |
226 | |
227 | As the analyzer analyzes a path, it may accrue information to refine the |
228 | knowledge about the type of an object. This can then be used to make better |
229 | decisions about the target method of a call. |
230 | |
231 | Such type information is tracked as DynamicTypeInfo. This is path-sensitive |
232 | data that is stored in ProgramState, which defines a mapping from MemRegions to |
233 | an (optional) DynamicTypeInfo. |
234 | |
235 | If no DynamicTypeInfo has been explicitly set for a MemRegion, it will be lazily |
236 | inferred from the region's type or associated symbol. Information from symbolic |
237 | regions is weaker than from true typed regions. |
238 | |
239 | EXAMPLE: A C++ object declared "A obj" is known to have the class 'A', but a |
240 | reference "A &ref" may dynamically be a subclass of 'A'. |
241 | |
242 | The DynamicTypePropagation checker gathers and propagates DynamicTypeInfo, |
243 | updating it as information is observed along a path that can refine that type |
244 | information for a region. |
245 | |
246 | WARNING: Not all of the existing analyzer code has been retrofitted to use |
247 | DynamicTypeInfo, nor is it universally appropriate. In particular, |
248 | DynamicTypeInfo always applies to a region with all casts stripped |
249 | off, but sometimes the information provided by casts can be useful. |
250 | |
251 | |
252 | RuntimeDefinition |
253 | ^^^^^^^^^^^^^^^^^ |
254 | |
255 | The basis of devirtualization is CallEvent's getRuntimeDefinition() method, |
256 | which returns a RuntimeDefinition object. When asked to provide a definition, |
257 | the CallEvents for dynamic calls will use the DynamicTypeInfo in their |
258 | ProgramState to attempt to devirtualize the call. In the case of no dynamic |
259 | dispatch, or perfectly constrained devirtualization, the resulting |
260 | RuntimeDefinition contains a Decl corresponding to the definition of the called |
261 | function, and RuntimeDefinition::mayHaveOtherDefinitions will return FALSE. |
262 | |
263 | In the case of dynamic dispatch where our information is not perfect, CallEvent |
264 | can make a guess, but RuntimeDefinition::mayHaveOtherDefinitions will return |
265 | TRUE. The RuntimeDefinition object will then also include a MemRegion |
266 | corresponding to the object being called (i.e., the "receiver" in Objective-C |
267 | parlance), which ExprEngine uses to decide whether or not the call should be |
268 | inlined. |
269 | |
270 | Inlining Dynamic Calls |
271 | ^^^^^^^^^^^^^^^^^^^^^^ |
272 | |
273 | The -analyzer-config ipa option has five different modes: none, basic-inlining, |
274 | inlining, dynamic, and dynamic-bifurcate. Under -analyzer-config ipa=dynamic, |
275 | all dynamic calls are inlined, whether we are certain or not that this will |
276 | actually be the definition used at runtime. Under -analyzer-config ipa=inlining, |
277 | only "near-perfect" devirtualized calls are inlined*, and other dynamic calls |
278 | are evaluated conservatively (as if no definition were available). |
279 | |
280 | * Currently, no Objective-C messages are not inlined under |
281 | -analyzer-config ipa=inlining, even if we are reasonably confident of the type |
282 | of the receiver. We plan to enable this once we have tested our heuristics |
283 | more thoroughly. |
284 | |
285 | The last option, -analyzer-config ipa=dynamic-bifurcate, behaves similarly to |
286 | "dynamic", but performs a conservative invalidation in the general virtual case |
287 | in *addition* to inlining. The details of this are discussed below. |
288 | |
289 | As stated above, -analyzer-config ipa=basic-inlining does not inline any C++ |
290 | member functions or Objective-C method calls, even if they are non-virtual or |
291 | can be safely devirtualized. |
292 | |
293 | |
294 | Bifurcation |
295 | ^^^^^^^^^^^ |
296 | |
297 | ExprEngine::BifurcateCall implements the ``-analyzer-config ipa=dynamic-bifurcate`` |
298 | mode. |
299 | |
300 | When a call is made on an object with imprecise dynamic type information |
301 | (RuntimeDefinition::mayHaveOtherDefinitions() evaluates to TRUE), ExprEngine |
302 | bifurcates the path and marks the object's region (retrieved from the |
303 | RuntimeDefinition object) with a path-sensitive "mode" in the ProgramState. |
304 | |
305 | Currently, there are 2 modes: |
306 | |
307 | * ``DynamicDispatchModeInlined`` - Models the case where the dynamic type information |
308 | of the receiver (MemoryRegion) is assumed to be perfectly constrained so |
309 | that a given definition of a method is expected to be the code actually |
310 | called. When this mode is set, ExprEngine uses the Decl from |
311 | RuntimeDefinition to inline any dynamically dispatched call sent to this |
312 | receiver because the function definition is considered to be fully resolved. |
313 | |
314 | * ``DynamicDispatchModeConservative`` - Models the case where the dynamic type |
315 | information is assumed to be incorrect, for example, implies that the method |
316 | definition is overridden in a subclass. In such cases, ExprEngine does not |
317 | inline the methods sent to the receiver (MemoryRegion), even if a candidate |
318 | definition is available. This mode is conservative about simulating the |
319 | effects of a call. |
320 | |
321 | Going forward along the symbolic execution path, ExprEngine consults the mode |
322 | of the receiver's MemRegion to make decisions on whether the calls should be |
323 | inlined or not, which ensures that there is at most one split per region. |
324 | |
325 | At a high level, "bifurcation mode" allows for increased semantic coverage in |
326 | cases where the parent method contains code which is only executed when the |
327 | class is subclassed. The disadvantages of this mode are a (considerable?) |
328 | performance hit and the possibility of false positives on the path where the |
329 | conservative mode is used. |
330 | |
331 | Objective-C Message Heuristics |
332 | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ |
333 | |
334 | ExprEngine relies on a set of heuristics to partition the set of Objective-C |
335 | method calls into those that require bifurcation and those that do not. Below |
336 | are the cases when the DynamicTypeInfo of the object is considered precise |
337 | (cannot be a subclass): |
338 | |
339 | - If the object was created with +alloc or +new and initialized with an -init |
340 | method. |
341 | |
342 | - If the calls are property accesses using dot syntax. This is based on the |
343 | assumption that children rarely override properties, or do so in an |
344 | essentially compatible way. |
345 | |
346 | - If the class interface is declared inside the main source file. In this case |
347 | it is unlikely that it will be subclassed. |
348 | |
349 | - If the method is not declared outside of main source file, either by the |
350 | receiver's class or by any superclasses. |
351 | |
352 | C++ Caveats |
353 | ^^^^^^^^^^^ |
354 | |
355 | C++11 [class.cdtor]p4 describes how the vtable of an object is modified as it is |
356 | being constructed or destructed; that is, the type of the object depends on |
357 | which base constructors have been completed. This is tracked using |
358 | DynamicTypeInfo in the DynamicTypePropagation checker. |
359 | |
360 | There are several limitations in the current implementation: |
361 | |
362 | * Temporaries are poorly modeled right now because we're not confident in the |
363 | placement of their destructors in the CFG. We currently won't inline their |
364 | constructors unless the destructor is trivial, and don't process their |
365 | destructors at all, not even to invalidate the region. |
366 | |
367 | * 'new' is poorly modeled due to some nasty CFG/design issues. This is tracked |
368 | in PR12014. 'delete' is not modeled at all. |
369 | |
370 | * Arrays of objects are modeled very poorly right now. ExprEngine currently |
371 | only simulates the first constructor and first destructor. Because of this, |
372 | ExprEngine does not inline any constructors or destructors for arrays. |
373 | |
374 | |
375 | CallEvent |
376 | ^^^^^^^^^ |
377 | |
378 | A CallEvent represents a specific call to a function, method, or other body of |
379 | code. It is path-sensitive, containing both the current state (ProgramStateRef) |
380 | and stack space (LocationContext), and provides uniform access to the argument |
381 | values and return type of a call, no matter how the call is written in the |
382 | source or what sort of code body is being invoked. |
383 | |
384 | NOTE: For those familiar with Cocoa, CallEvent is roughly equivalent to |
385 | NSInvocation. |
386 | |
387 | CallEvent should be used whenever there is logic dealing with function calls |
388 | that does not care how the call occurred. |
389 | |
390 | Examples include checking that arguments satisfy preconditions (such as |
391 | __attribute__((nonnull))), and attempting to inline a call. |
392 | |
393 | CallEvents are reference-counted objects managed by a CallEventManager. While |
394 | there is no inherent issue with persisting them (say, in a ProgramState's GDM), |
395 | they are intended for short-lived use, and can be recreated from CFGElements or |
396 | non-top-level StackFrameContexts fairly easily. |
397 | |